data: add benchmark result lemer-bf16-math-temp0.json
Browse files
results/lemer-bf16-math-temp0.json
ADDED
|
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model": "/Users/snider/.cache/huggingface/hub/models--lthn--lemer/snapshots/083106e3de02de3606452d58fa0ee2fcb2faff92",
|
| 3 |
+
"benchmark": "MMLU-Pro",
|
| 4 |
+
"method": "rapid-mlx + openai SDK + google parse_response",
|
| 5 |
+
"temperature": 0.0,
|
| 6 |
+
"processor": "google/gemma-4-E2B-it",
|
| 7 |
+
"num_runs": 1,
|
| 8 |
+
"limit_per_subject": 20,
|
| 9 |
+
"runs": [
|
| 10 |
+
{
|
| 11 |
+
"math": {
|
| 12 |
+
"correct": 11,
|
| 13 |
+
"total": 20,
|
| 14 |
+
"accuracy": 0.55
|
| 15 |
+
},
|
| 16 |
+
"_overall": {
|
| 17 |
+
"correct": 11,
|
| 18 |
+
"total": 20,
|
| 19 |
+
"accuracy": 0.55
|
| 20 |
+
},
|
| 21 |
+
"_elapsed_s": 345.0
|
| 22 |
+
}
|
| 23 |
+
]
|
| 24 |
+
}
|